How EvalCite Works
EvalCite verifies each reference against 8 authoritative academic databases and tells you whether the cited paper actually exists with the details you gave. Here's the whole pipeline, step by step.
Step 1 — Parse the reference
Your reference list is split into individual entries and each is parsed into structured fields: title, authors, year, venue, DOI, arXiv ID. This works across the formats reference lists actually come in — numbered IEEE-style lists wrapped across lines, APA/Chicago author-year lists, and BibTeX files. PDFs are converted to text first, with two-column reading order handled.
Example: Smith, J. et al. (2020). Machine Learning Basics. Nature, 50(1). becomes title = "Machine Learning Basics", authors = "Smith, J.; et al.", year = 2020, venue = "Nature".
Step 2 — Query 8 academic databases
Every reference is searched across multiple authoritative sources simultaneously:
| Database | Covers |
|---|---|
| CrossRef | The DOI registry — journals, conference papers, books with DOIs |
| OpenAlex | 250M+ scholarly works across all fields |
| arXiv | Preprints — physics, CS, math, and more |
| Semantic Scholar | Papers with strong abstract coverage |
| DBLP | Computer-science conferences and journals |
| OpenLibrary | 1.7M+ books |
| IEEE Xplore | Engineering & CS (when an API key is configured) |
| Scopus | Premium citation database (when an API key is configured) |
Step 3 — Match the real paper
Candidate records from the databases are compared against your reference on title and authors, with a year check on top. An exact DOI or arXiv ID resolves directly to the authoritative record and verifies instantly. When a reference carries a link no database knows — a blog post, a model card, documentation — the live page itself is fetched and its title compared, so real web sources aren't wrongly flagged.
Step 4 — Return your verdict
Each reference gets one of three verdicts:
- Verified — a database record matches. The paper exists with the cited details.
- Mismatch — a similar paper was found but at least one field disagrees; you see exactly which field and what the authoritative record says.
- Not found — nothing close exists in any database. For AI-generated text, this is the classic hallucination signal.
You also get a confidence score per reference, a link to the matched record, and a downloadable corrected .bib file: verified entries are rebuilt from the authoritative metadata, unverifiable ones kept as parsed and clearly marked.
Why you can trust it
- Transparent by design — every verdict links to the exact database record that confirmed (or failed to confirm) the paper.
- Exact identifiers verify instantly — a matching DOI or arXiv ID is conclusive.
- Known limitations are published — see the accuracy & methodology report for where the system is weak (very recent papers, non-scholarly sources, rate limits).
- Your documents stay private — PDFs are processed in memory and deleted immediately after the references are extracted. See the privacy policy.